高分辨率遥感影像的随机森林变化检测方法

发布时间：2018-08-28 16:53

【摘要】：基于面向对象分析(OBIA)的遥感影像变化检测研究已取得显著的进展,代表了遥感影像变化检测的发展范式,未来是发展更加智能的解译分析方法。随机森林作为一种新的机器学习算法,其预测效果和性能稳定性要优于许多单预测器和集成预测方法。本文充分利用OBIA及随机森林机器学习算法的优势,提出了利用随机森林进行面向对象的遥感影像变化检测。首先基于熵率对影像进行超像素分割,通过最优超像素个数评价指数来获取最佳的影像分割结果,并提取每个超像素在前、后时相影像上的光谱特征和Gabor特征作为随机森林的特征输入数据,用于模型的训练。在初始像素级检测结果之上,自动进行分类样本选择并构建分类器模型,用训练好的模型来提取最终的变化区域。利用Quickbird、IKONOS、SPOT-5等3组多光谱影像进行试验,结果表明,本文方法在变化检测精度上要优于对比方法。
[Abstract]:The research of remote sensing image change detection based on object oriented analysis (OBIA) has made remarkable progress, which represents the development paradigm of remote sensing image change detection, and will be a more intelligent interpretation and analysis method in the future. As a new machine learning algorithm, stochastic forest has better prediction effect and performance stability than many single predictors and integrated prediction methods. Based on the advantages of OBIA and stochastic forest machine learning algorithm, an object oriented remote sensing image change detection based on stochastic forest is proposed in this paper. Firstly, the image is segmented based on entropy rate, and the best image segmentation results are obtained by evaluating the optimal number of super-pixels, and each super-pixel is extracted before the image segmentation. Spectral features and Gabor features in post-temporal images are used as input data for model training. Based on the initial pixel level detection results, the classifier model is automatically selected and the trained model is used to extract the final change region. Three groups of multispectral images, such as Quickbird,IKONOS,SPOT-5, are used for experiments. The results show that the proposed method is superior to the contrast method in the accuracy of change detection.
【作者单位】：武汉大学测绘遥感信息工程国家重点实验室;长江大学电子信息学院;隆德大学自然地理和生态系统科学系;
【基金】：国家重点研发计划(2016YFB0502603) 测绘遥感信息工程国家重点实验室开放基金(16E01) 国家自然科学基金(41471354)~~
【分类号】：TP751

【相似文献】